Practical overview
When it fits
Use ReQFlow when a research workflow needs configurable, local inference for protein backbone generation and access to the accompanying model code and checkpoints.
Who it is for
- Protein-design researchers generating candidate backbone structures.
- Researchers studying flow-based generative methods for protein structures.
- Developers exploring point-cloud and protein-backbone generation workflows.
Documented uses
- Generate unconditional candidate protein backbones and save structures as PDB files.
- Configure sampling by backbone length, sample count and inference settings.
- Use ProteinMPNN and ESMFold in the documented self-consistency evaluation workflow.
- Inspect generated results for downstream designability, diversity and novelty analysis.
Requirements & limits
The repository links model weights on Hugging Face and Google Drive and documents local installation and command-line inference.
- The documented primary task is unconditional backbone generation, configured through YAML and launched from the command line.
- The example inference configuration specifies four GPUs and recommends lengths below 500.
- The described evaluation workflow uses ProteinMPNN and ESMFold in addition to ReQFlow.
- The README reports speed comparisons against other models; these are project claims rather than independently verified results.
Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.
Set up ReQFlow for unconditional inference
Install the repository environment, prepare a checkpoint and configure the documented inference workflow.
Before you start
- The ReQFlow repository and a conda or mamba environment manager.
- A ReQFlow checkpoint from the linked Hugging Face or Google Drive location, or user-provided weights.
- The README’s listed torch-scatter and DeepSpeed dependencies.
Steps
Create the environment from reqflow-env.yml and activate reqflow-env.
Install torch-scatter using the README’s PyTorch 2.0.0 with CUDA 11.7 wheel index, then upgrade DeepSpeed.
From the ReQFlow directory, install the local package with pip install -e .
Place checkpoint and configuration files in the documented ckpts layout, then edit configs/inference_unconditional.yaml for the checkpoint path and sampling settings.
Expected outcome
The repository and configuration are prepared for the documented inference command; this setup guide does not claim that inference has been run.
Guide limitations
- The example configuration specifies four GPUs and refers to ProteinMPNN and ESMFold paths for evaluation.
- The README recommends a maximum sampling length below 500.
Configure and launch unconditional backbone inference
Use the repository’s documented configuration and command to request unconditional protein backbone samples.
Before you start
- A prepared ReQFlow environment and locally installed package.
- A checkpoint with its configuration, referenced by ckpt_path.
- An edited configs/inference_unconditional.yaml file.
Steps
Set the inference task to unconditional and provide the checkpoint and output-directory paths.
Choose sampling lengths, length increment, samples per length and the other inference settings in the YAML configuration.
Run python -W ignore experiments/inference_se3_flows.py -cn inference_unconditional from the repository.
Expected outcome
The README describes outputs organized by run and sampled length, including final PDB samples and an inference configuration; evaluation folders are also described for the ProteinMPNN and ESMFold workflow.
Guide limitations
- The example uses four GPUs and recommends lengths below 500.
- The repository links to checkpoint downloads, but this guide does not verify download or runtime success.
Sources & review
Source record details
- Source availability
- Active
- Sources checked
- 18 October 2026
- Original website title
- AngxiaoYue/ReQFlow · Official homepage title · retained in its source language